Kindred passages

What consumers told the committee, and where else in the record they said it.

Since 2014 each Public Summary Document has carried the PBAC's own summary of the comments it received from patients, carers, clinicians and organisations. Using an SBERT semantic-embedding approach adapted from peer-reviewed work by Luc et al. (2026), this page finds passages elsewhere in the record that read most like each summary, chunk against chunk, across every medicine. The text is the committee's published summary, verbatim. Nothing is generated on this page.

Source documents

805

Random-pair baseline

mean 0.34 · p95 0.56 · p99 0.71

Top passage beats p99

693 of 805

Same medicine hidden

567 of 768

Method: SBERT chunk embeddings, 768-d, ~100-word chunks with 25-word overlap (Isaiah, delivered 2026-09-18). Chunks led by the standard opening line and "no consumer comments" sections are excluded, since the template dominates the vector whatever follows it. 1,964 of 4,109 chunks kept; 1,802 of the embedded set's filenames join the script.report record. Baseline is the cosine between 39,874 random pairs of chunks from different documents. Built 2026-09-29.